During February 2025, the developer contributed to the fzyzcjy/sglang repository by implementing block-wise INT8 quantization support for DeepSeek V3/R1 models. Leveraging C++, CUDA, and Python, they introduced new quantization methods and custom kernels to optimize deep learning model inference. Their work focused on improving inference throughput and reducing deployment costs by enabling more efficient model execution. To ensure reliability, they developed a comprehensive test suite that validated both accuracy and performance gains. The depth of the contribution is reflected in the integration of model optimization techniques and thorough validation, enhancing the deployment readiness of quantized models in production environments.